Physicists have developed a deep understanding of the fundamental laws of nature through careful observations, experiments, and precise measurements. However, what if artificial intelligence (AI) could uncover governing laws of the physical world by analyzing data on its own?
Researchers at Archetype AI have developed a foundational AI model, called the “Newton AI Model”, capable of learning complex physics principles directly from raw sensor data. This potentially game-changing model can accurately predict various physical phenomena, including those never encountered during training.
Unlike the typical AI models, which rely on pre-programmed knowledge from datasets, Newton AI relies on analyzing sensor measurements, such as temperature, vibrations, and pressure, to build its own understanding of how physical systems work.
Newton AI represents a significant leap forward for the use of AI in physics, but it’s not just limited to science. Newton AI can learn and predict behaviors across a range of environments and systems that it wasn't explicitly trained on.
Newton AI's ability to accurately predict data it has not previously encountered is known as zero-shot learning. While several other models demonstrate similar zero-shot learning capabilities, what sets Newton AI apart is its specific application to physics and complex physical systems.
Archetype AI claims that Newton AI outperforms specialized AI models in real-world physical applications such as forecasting city-wide power consumption or predicting the oil temperature of an electrical transformer.
When you train an AI foundation model on fundamental laws of physics, such as the conservation of energy, you may introduce "inductive biases." These biases can limit the model to existing knowledge and understanding, potentially hindering its ability to draw new conclusions from the data.
According to Archetype AI, the approach taken with Newton AI of discovering new physical phenomena directly from data is how humans have historically uncovered the fundamental laws of nature.
Some of the most important physical laws, such as laws of electricity or planetary motion, were discovered without prior knowledge of the underlying principles. Scientists observed, measured, and analyzed physical systems to reveal patterns, which we now recognize as the laws of nature.
The team at Archetype AI is aware of the challenges of applying GenAI principles to the complex realm of physical data and predicting the future behavior of diverse physical systems. To address this challenge, Archetype AI has given Newton AI dual capabilities of being both pre-trained and capable of zero-shot learning.
Newton AI is pre-trained on 590 million samples for open-source datasets. Utilizing a transformer-based deep neural network, Newton analyzes raw sensor data to uncover hidden patterns. This allows Newton AI to process both real-time and recorded sensor data effectively.
A blog by the Archetype AI team highlighted, “What’s exciting is that zero-shot forecasting consistently outperforms even when Newton is specifically trained on only the target dataset. In other words, Newton, trained on a vast amount of physical sensor data from all over the world, knows more about the temperature of oil in transformers than when we train Newton specifically on transformer oil data.”
“This suggests that foundation models like Newton have powerful generalization capabilities, enabling them to understand physical behaviors far beyond the specific data they were originally trained on.”
Newton AI's ability to generalize could be particularly advantageous in situations where data is scarce or challenging to collect. Using its broad base of physical knowledge, Newton AI could provide valuable insights in various contexts where data is limited.
The capabilities offered by Newton AI could also enhance human perception, enabling the detection of those aspects of our physical world that are imperceptible to us. This holds promise for driving breakthroughs in fields such as medicine.
Archetype AI, a Palo Alto-based startup founded by former Google researchers, aims to leverage Newton AI to gain a deeper understanding of our physical world. Currently, Newton AI is only a research prototype, but with further development, it has the potential to be brought to market, ushering in a new era of AI-driven insights into our physical environment.
